In wireless sensor networks, link metric estimation at each hop should not require a long history of packet exchanges. In this paper, we explore several approaches to link quality estimation. We report on the results of experiments on the Grenoble testbed of the FIT IoT-lab composed of a set of Cortex M3 nodes with IEEE 802.15.4 radios. Whereas the received signal power is a poor indication of PDR (Packet Delivery Ratio) that one can expect on a given link, LQI (Link Quality Indicator) gives more accurate information. We propose a two stage classification, in which a very large fraction of links are immediately either deemed usable or not, while the remaining ones need a bit more testing before they are advertised by the routing protocol as good or weak links.